orchestration trained

the learned layer: read first.

mental model

Delegate only when authorized and a task has an independent boundary. Give agents disjoint writes, one integration owner, and explicit evidence checks.

examples

outcome: inspect the routing contract
read scope: owned spec and router
write scope: none
completion check: cite exact lines and unresolved gaps
model: gpt-6-luna, explicit reasoning effort

best practices

Specify outcome, evidence, exact scope, dependencies, and constraints.
Verify available model identifiers before delegation.
Reconcile agent reports with fresh source; lead runs final checks.

strengths

Independent inventory, implementation, and review tasks can advance while one owner protects integration.

weaknesses / pain points

Shared mutable files defeat parallel speed. Serialize overlapping edits.

gotchas

Imported team protocols may assume worktrees, branches, or tool names not present in this runtime. Those require explicit authority or verification.

sources and conflicts

The upstream subagent workflow informs task review. Official Claude team documentation describes an optional, runtime-specific team mechanism. Neither source grants delegation or worktree authority in this workspace; the operator's instructions require explicit delegation scope, a shared checkout, and one integration owner.

known bugs

No version-specific agent runtime defect has been verified here.

troubleshooting

Conflicting agent edits resolve by assigning disjoint write paths or serializing the mutation. If the requested model name does not exist in the runtime, read the current model list before dispatch and use the closest authorized tier.

Read the orchestration skill.